Fix cross-page PDF QA recall and add regression coverage

[AkilsuryaS]

Summary

• Use vdb_top_k as the retrieval candidate pool in the RAG generation path to improve recall when answers require evidence across multiple PDF pages.
• Update RAG answer prompts to explicitly allow synthesis across multiple context chunks.
• Add regression coverage for vdb_top_k behavior when reranker is disabled, and tighten the integration check for search result count mismatches.

Issue linkage

Closes #240

Test plan

• python -m py_compile src/nvidia_rag/rag_server/main.py tests/unit/test_rag_server/test_query_rewriting.py tests/integration/test_cases/rag_search.py
• Run unit/integration suite in project environment with pinned dependencies.

Made with Cursor

[JiwaniZakir]

The removal of the branches filter from the pull_request trigger in ci-pipeline.yml (lines 7-8) is a significant behavioral change — the pipeline will now fire on PRs targeting any branch, not just develop. Combined with adding main and release-** to the push trigger, this could produce duplicate runs for release branches and substantially increase CI load; the comment says "avoids duplicate runs when pushing to a PR branch" but the missing branch filter does the opposite for PRs.

In the Prepare multimodal test data step, the [ -f ... ] && cp ... || true guards silently succeed when none of the source files exist. If the test fixtures are absent, the multimodal query tests will fail later with a confusing error rather than a clear "missing test data" signal at preparation time — it would be worth adding an explicit check or echo so failures surface at the right step.

The sleep 5 before bringing Docker services back up in Restart services for multimodal query tests is inconsistent with the sleep 30 used in every other restart block in this file; if services aren't ready before the test step runs, the multimodal tests will fail non-deterministically in CI even though the continue-on-error: true will mask it.